WEBVTT

1
00:00:00.103 --> 00:00:02.552
Imagine asking an AI to build a family

2
00:00:02.564 --> 00:00:04.987
meal planner It produces a polished app.

3
00:00:05.425 --> 00:00:08.080
The buttons work the tests pass and the

4
00:00:08.093 --> 00:00:10.954
calendar fills with dinners But your family still

5
00:00:10.966 --> 00:00:12.100
argues about food

6
00:00:13.196 --> 00:00:15.941
The software runs Yet the experience you wanted

7
00:00:15.954 --> 00:00:19.807
never arrived Daniel Miessler's article Writing Is Closer

8
00:00:19.820 --> 00:00:22.474
to Thought Than Code asks us to examine

9
00:00:22.487 --> 00:00:24.949
the gap between the idea in our heads

10
00:00:25.155 --> 00:00:27.307
and the instructions we actually give

11
00:00:28.383 --> 00:00:30.651
In this lesson you will learn his argument

12
00:00:30.896 --> 00:00:33.705
build a small Ideal State Artifact and connect

13
00:00:33.718 --> 00:00:36.927
its criteria to evaluations The aim is not

14
00:00:36.940 --> 00:00:39.646
better sounding prompts It is a clearer target

15
00:00:39.659 --> 00:00:40.290
for the work

16
00:00:41.343 --> 00:00:44.024
Miessler frames his argument as a disagreement with

17
00:00:44.036 --> 00:00:47.194
developer Matt Pocock In the position he quotes

18
00:00:47.426 --> 00:00:50.106
code belongs at the center because code is

19
00:00:50.119 --> 00:00:54.294
executable Specifications help but the implementation is the

20
00:00:54.307 --> 00:00:55.286
main artifact

21
00:00:56.343 --> 00:00:59.371
Miessler instead puts articulated intent at the center

22
00:00:59.835 --> 00:01:02.245
what are we building for whom and why

23
00:01:02.683 --> 00:01:04.539
His claim is that this is a higher

24
00:01:04.552 --> 00:01:07.065
fidelity record of the original idea than the

25
00:01:07.078 --> 00:01:09.217
code that happens to implement it today

26
00:01:10.250 --> 00:01:12.969
A useful distinction is this code tells us

27
00:01:12.982 --> 00:01:15.998
what the system actually does Written intent tells

28
00:01:16.010 --> 00:01:18.163
us what we want it to accomplish The

29
00:01:18.175 --> 00:01:20.972
article argues that the second question should direct

30
00:01:20.985 --> 00:01:24.980
the first especially when AI does the implementation

31
00:01:26.076 --> 00:01:29.633
Picture three stages First comes thought a parent

32
00:01:29.710 --> 00:01:33.125
imagining a calmer Sunday evening That intention contains

33
00:01:33.138 --> 00:01:37.661
context feelings priorities and tradeoffs None of those

34
00:01:37.674 --> 00:01:39.310
automatically reach the AI

35
00:01:40.383 --> 00:01:44.584
Second comes articulation Writing and sometimes images make

36
00:01:44.597 --> 00:01:48.257
the intention available outside your head Miessler treats

37
00:01:48.270 --> 00:01:50.924
writing as closest to thought In his model

38
00:01:51.311 --> 00:01:54.378
it is the first representation code is another

39
00:01:54.391 --> 00:01:55.241
step removed

40
00:01:56.368 --> 00:01:59.898
Third comes implementation The same desired experience could

41
00:01:59.911 --> 00:02:04.666
produce different databases interfaces or programming languages Those

42
00:02:04.679 --> 00:02:07.308
choices matter but they are means If the

43
00:02:07.321 --> 00:02:10.337
means change we still need a stable explanation

44
00:02:10.350 --> 00:02:10.942
of the end

45
00:02:12.023 --> 00:02:15.013
Why is articulation difficult The article points to

46
00:02:15.026 --> 00:02:18.196
writer's blindness also called the curse of knowledge

47
00:02:18.621 --> 00:02:20.825
Once you know the context it becomes hard

48
00:02:20.838 --> 00:02:22.719
to notice what a reader does not know

49
00:02:23.796 --> 00:02:26.464
You think obviously the meal planner must avoid

50
00:02:26.476 --> 00:02:29.440
peanuts But your prompt only says make a

51
00:02:29.453 --> 00:02:32.778
weekly meal planner The missing allergy requirement feels

52
00:02:32.791 --> 00:02:35.291
obvious to you because you already carry it

53
00:02:35.304 --> 00:02:36.296
in your mental model

54
00:02:37.343 --> 00:02:40.449
Miessler argues that many apparent AI failures are

55
00:02:40.462 --> 00:02:44.250
actually communication failures Treat that as a diagnostic

56
00:02:44.263 --> 00:02:47.859
question not a universal excuse was the requirement

57
00:02:47.871 --> 00:02:52.395
absent misunderstood or implemented incorrectly Those are different

58
00:02:52.408 --> 00:02:53.993
failures with different fixes

59
00:02:55.076 --> 00:02:57.937
The article uses archery to explain the distinction

60
00:02:58.362 --> 00:03:01.725
Accuracy is closeness to the bullseye Here the

61
00:03:01.738 --> 00:03:04.432
bullseye means the outcome the person actually intended

62
00:03:04.857 --> 00:03:07.383
Shots can be near that target without landing

63
00:03:07.486 --> 00:03:09.071
in exactly the same place

64
00:03:10.156 --> 00:03:12.862
Precision is a tight grouping A program can

65
00:03:12.875 --> 00:03:15.723
implement a very exact rule while that rule

66
00:03:15.968 --> 00:03:18.275
serves the wrong goal Think of a planner

67
00:03:18.287 --> 00:03:21.548
that perfectly minimizes calories when the family really

68
00:03:21.561 --> 00:03:23.171
wanted less stress at dinner

69
00:03:24.223 --> 00:03:27.612
Miessler associates writing with accuracy and code with

70
00:03:27.625 --> 00:03:30.834
precision This is an analogy not a guarantee

71
00:03:30.847 --> 00:03:33.231
about either medium Vague prose can miss the

72
00:03:33.244 --> 00:03:36.362
goal too The practical lesson is to establish

73
00:03:36.375 --> 00:03:39.713
the right target then demand reliable implementation

74
00:03:40.783 --> 00:03:43.837
Miessler's proposed working document is an Ideal State

75
00:03:43.850 --> 00:03:46.569
Artifact or ISA It is more than a

76
00:03:46.582 --> 00:03:49.855
rigid specification It begins with prose that captures

77
00:03:49.868 --> 00:03:52.317
the desired state as fully as the problem

78
00:03:52.330 --> 00:03:52.832
needs

79
00:03:53.863 --> 00:03:56.569
It also records constraints things we must not

80
00:03:56.582 --> 00:04:00.306
do And it records decisions including choices made

81
00:04:00.319 --> 00:04:03.605
while the goal becomes clearer The article explicitly

82
00:04:03.683 --> 00:04:06.724
allows understanding to evolve as we build You

83
00:04:06.737 --> 00:04:08.335
do not need perfect foresight

84
00:04:09.513 --> 00:04:14.204
Finally come Ideal State Criteria or ISCs atomic

85
00:04:14.410 --> 00:04:17.490
testable pieces of functionality each paired with an

86
00:04:17.503 --> 00:04:21.614
evaluation The artifact holds the vision its boundaries

87
00:04:21.781 --> 00:04:23.869
and the checks that connect that vision to

88
00:04:23.882 --> 00:04:25.287
a real implementation

89
00:04:26.343 --> 00:04:29.436
Now apply the article's example The ideal state

90
00:04:29.500 --> 00:04:31.884
is a Sunday evening where the family chooses

91
00:04:31.897 --> 00:04:34.165
a week of dinners in under five minutes

92
00:04:34.487 --> 00:04:36.652
everyone is excited to eat them and the

93
00:04:36.665 --> 00:04:38.431
groceries are already in the cart

94
00:04:39.503 --> 00:04:40.972
Notice what this gives us that build a

95
00:04:40.985 --> 00:04:43.627
meal planner does not We have a person

96
00:04:44.026 --> 00:04:47.287
a situation a time budget an emotional outcome

97
00:04:47.583 --> 00:04:50.444
and a completed task A calendar screen alone

98
00:04:50.508 --> 00:04:52.403
does not satisfy that description

99
00:04:53.463 --> 00:04:56.195
As a practical extension ask the AI to

100
00:04:56.221 --> 00:04:59.082
interview you about gaps What does excited mean

101
00:04:59.095 --> 00:05:02.265
for this family Which grocery service Who approves

102
00:05:02.278 --> 00:05:05.847
substitutions The purpose of discussion is to expose

103
00:05:05.860 --> 00:05:09.159
missing context not decorate the document with extra

104
00:05:09.172 --> 00:05:09.649
words

105
00:05:10.703 --> 00:05:13.925
The example's first constraint is no peanuts because

106
00:05:13.938 --> 00:05:16.670
Maya has an allergy Its second is that

107
00:05:16.682 --> 00:05:20.020
weeknight dinners take forty minutes or less These

108
00:05:20.033 --> 00:05:22.739
are boundaries that candidate plans must respect

109
00:05:23.796 --> 00:05:26.747
The decisions section explains why calorie tracking was

110
00:05:26.760 --> 00:05:29.337
dropped it made planning feel like a chore

111
00:05:29.775 --> 00:05:33.229
That rationale matters Otherwise an AI or a

112
00:05:33.242 --> 00:05:37.018
future collaborator might reintroduce the feature believing more

113
00:05:37.031 --> 00:05:38.796
functionality must be better

114
00:05:39.863 --> 00:05:42.724
A safety caveat for this teaching example checking

115
00:05:42.737 --> 00:05:45.623
ingredient names is not a complete allergy safety

116
00:05:45.636 --> 00:05:50.250
system Labels substitutions and cross contact require attention

117
00:05:50.263 --> 00:05:53.407
too Use the example to understand the structure

118
00:05:53.832 --> 00:05:56.397
not as clinical guidance or a production safety

119
00:05:56.410 --> 00:05:56.848
design

120
00:05:57.903 --> 00:06:00.609
The article breaks the planner into five criteria

121
00:06:01.035 --> 00:06:04.102
No planned dish contains peanuts The grocery list

122
00:06:04.114 --> 00:06:07.761
covers every ingredient Every weeknight dinner takes forty

123
00:06:07.774 --> 00:06:10.455
minutes or less These are the first three

124
00:06:10.468 --> 00:06:11.009
checks

125
00:06:12.076 --> 00:06:13.893
The fourth asks whether a ten year old

126
00:06:13.906 --> 00:06:16.058
would want to eat each dinner The fifth

127
00:06:16.071 --> 00:06:18.506
asks whether the week feels varied rather than

128
00:06:18.519 --> 00:06:21.084
repetitive Now we can see a difference between

129
00:06:21.097 --> 00:06:23.983
mechanical checks and judgments about the experience

130
00:06:25.036 --> 00:06:27.948
Keep each criterion narrow enough that failure tells

131
00:06:27.961 --> 00:06:31.428
you something useful Grocery completeness and dinner variety

132
00:06:31.647 --> 00:06:34.301
should not share one pass or fail box

133
00:06:34.740 --> 00:06:37.665
They can fail independently and the system needs

134
00:06:37.678 --> 00:06:39.946
different changes to improve each one

135
00:06:41.048 --> 00:06:44.166
For a deterministic claim use a deterministic check

136
00:06:44.256 --> 00:06:47.310
when the data supports it Compare required ingredients

137
00:06:47.323 --> 00:06:50.597
with the grocery list Check recorded preparation times

138
00:06:50.609 --> 00:06:53.496
against the limit An assertion is only as

139
00:06:53.509 --> 00:06:56.202
meaningful as the data and assumptions behind it

140
00:06:57.302 --> 00:06:59.944
For kid appeal the article proposes a question

141
00:06:59.956 --> 00:07:03.603
set As an instructional example ask about familiar

142
00:07:03.616 --> 00:07:06.876
flavors difficult textures and whether the family has

143
00:07:06.889 --> 00:07:10.124
actually liked similar dishes A model's answer is

144
00:07:10.253 --> 00:07:13.204
evidence to inspect not the child's own preference

145
00:07:14.302 --> 00:07:16.699
For variety use a rubric or a model

146
00:07:16.712 --> 00:07:20.603
judge Define dimensions such as repeated main ingredients

147
00:07:20.835 --> 00:07:25.371
cooking methods and flavors Miessler describes evaluations ranging

148
00:07:25.384 --> 00:07:28.954
from basic assertions to multi stage processes and

149
00:07:28.967 --> 00:07:32.794
model judges Choose the simplest evaluation that really

150
00:07:32.807 --> 00:07:33.735
tests the claim

151
00:07:34.809 --> 00:07:37.979
Here is the article's strongest structural idea The

152
00:07:37.992 --> 00:07:40.595
criteria tell the AI what to build and

153
00:07:40.608 --> 00:07:43.430
the evaluations tell us whether it succeeded They

154
00:07:43.443 --> 00:07:46.059
are not separate stories about the project They

155
00:07:46.072 --> 00:07:47.953
share one definition of success

156
00:07:49.036 --> 00:07:53.263
Take grocery completeness That criterion drives recipe expansion

157
00:07:53.366 --> 00:07:56.626
and list generation Its evaluation then compares the

158
00:07:56.639 --> 00:07:59.371
ingredients required by the plan with the list

159
00:07:59.384 --> 00:08:03.224
the system actually produced If garlic disappears during

160
00:08:03.237 --> 00:08:06.755
aggregation we have a specific failure to investigate

161
00:08:07.836 --> 00:08:10.568
This does not mean prose executes itself The

162
00:08:10.581 --> 00:08:12.952
checks still need to be implemented and run

163
00:08:13.377 --> 00:08:16.857
The value is traceability for each behavior show

164
00:08:16.870 --> 00:08:19.344
the intention it serves and the evidence that

165
00:08:19.357 --> 00:08:20.452
it meets that intention

166
00:08:21.503 --> 00:08:25.008
Now imagine all five example criteria pass but

167
00:08:25.021 --> 00:08:27.392
a parent needs twelve minutes to finish planning

168
00:08:27.818 --> 00:08:30.627
This is an instructional extension not an outcome

169
00:08:30.640 --> 00:08:33.049
reported in the article What did our checklist

170
00:08:33.062 --> 00:08:33.423
miss

171
00:08:34.463 --> 00:08:37.414
The ideal state promised under five minutes yet

172
00:08:37.427 --> 00:08:40.584
the five listed criteria never directly tested that

173
00:08:40.597 --> 00:08:43.729
promise This is why a passing checklist cannot

174
00:08:43.741 --> 00:08:46.151
prove that the entire vision has been captured

175
00:08:46.577 --> 00:08:48.664
Compare the criteria back to the prose

176
00:08:49.703 --> 00:08:52.615
Add a timed planning task under stated conditions

177
00:08:53.002 --> 00:08:56.842
observe representative users and investigate the friction If

178
00:08:56.855 --> 00:08:59.884
the goal itself changes update the ideal state

179
00:09:00.128 --> 00:09:03.015
and record the decision The document should learn

180
00:09:03.028 --> 00:09:05.644
from use rather than become a frozen contract

181
00:09:05.657 --> 00:09:06.378
with reality

182
00:09:07.423 --> 00:09:10.593
Miessler's broader prediction is that as AI becomes

183
00:09:10.606 --> 00:09:14.060
better at execution the scarce human contribution shifts

184
00:09:14.073 --> 00:09:16.727
toward telling it exactly what we want He

185
00:09:16.740 --> 00:09:20.310
argues against keeping humans in every implementation step

186
00:09:20.323 --> 00:09:20.993
by default

187
00:09:22.089 --> 00:09:25.633
He invokes Richard Sutton's Bitter Lesson general methods

188
00:09:25.646 --> 00:09:28.416
that benefit from more computation tend to beat

189
00:09:28.429 --> 00:09:32.759
approaches built around handcrafted cleverness Miessler applies that

190
00:09:32.772 --> 00:09:36.522
lesson to development favoring clear goals over detailed

191
00:09:36.535 --> 00:09:40.169
micromanagement That application is his argument not a

192
00:09:40.182 --> 00:09:42.501
proof that all supervision is wasteful

193
00:09:43.543 --> 00:09:46.455
Our practical caveat is to separate direction from

194
00:09:46.468 --> 00:09:50.244
permission An AI can choose implementation details without

195
00:09:50.257 --> 00:09:53.388
being allowed to spend money release sensitive data

196
00:09:53.775 --> 00:09:57.486
or deploy unsafe changes Clear intent should reduce

197
00:09:57.577 --> 00:10:01.339
unnecessary steering not remove accountability for high stakes

198
00:10:01.352 --> 00:10:01.958
actions

199
00:10:03.036 --> 00:10:05.098
Pause and apply this to a real project

200
00:10:05.536 --> 00:10:08.655
Without looking back write one paragraph describing the

201
00:10:08.667 --> 00:10:12.611
ideal experience Name the person the situation and

202
00:10:12.624 --> 00:10:15.227
the change they should notice Avoid starting with

203
00:10:15.240 --> 00:10:16.425
a technology stack

204
00:10:17.503 --> 00:10:20.145
Then add two constraints and one decision with

205
00:10:20.158 --> 00:10:24.024
its reason Write three narrow criteria Beside each

206
00:10:24.256 --> 00:10:26.975
specify how you would check it an assertion

207
00:10:27.387 --> 00:10:30.171
a real user task a question set or

208
00:10:30.184 --> 00:10:32.696
a rubric Pause the video now and try

209
00:10:32.709 --> 00:10:32.877
it

210
00:10:33.943 --> 00:10:37.023
When you return test your understanding If every

211
00:10:37.036 --> 00:10:40.154
criterion passes could the person still be disappointed

212
00:10:40.580 --> 00:10:43.943
If yes identify the missing expectation If your

213
00:10:43.956 --> 00:10:46.649
only evaluation is ask the AI whether it

214
00:10:46.662 --> 00:10:49.368
did a good job replace that with observable

215
00:10:49.381 --> 00:10:49.923
evidence

216
00:10:51.009 --> 00:10:53.676
Here is the lesson in one chain Articulate

217
00:10:53.728 --> 00:10:56.885
the desired state Make its boundaries and tradeoffs

218
00:10:56.898 --> 00:11:01.576
visible Translate it into atomic criteria with evaluations

219
00:11:02.001 --> 00:11:05.171
Build observe and revise what the evidence shows

220
00:11:05.184 --> 00:11:06.035
was incomplete

221
00:11:07.103 --> 00:11:09.552
Miessler's central claim is not that code has

222
00:11:09.564 --> 00:11:11.794
no value It is that code is a

223
00:11:11.807 --> 00:11:14.745
representation of a goal and written intent is

224
00:11:14.758 --> 00:11:17.876
closer to that goal The implementation should serve

225
00:11:17.889 --> 00:11:20.312
the idea not silently replace it

226
00:11:21.383 --> 00:11:24.875
This is an independent educational adaptation of Daniel

227
00:11:24.888 --> 00:11:28.561
Miessler's article published October twenty third twenty twenty

228
00:11:28.574 --> 00:11:32.788
six The narration is synthetic The diagnostic questions

229
00:11:32.801 --> 00:11:36.177
and safety caveats are teaching additions Before your

230
00:11:36.190 --> 00:11:39.025
next AI build ask have I described the

231
00:11:39.038 --> 00:11:41.460
target and what evidence would show we actually

232
00:11:41.473 --> 00:11:41.950
reached it